The Gold Standard of Geostatistics

Ordinary statistics can describe a set of values while ignoring where they were observed. Geostatistics adds location and models how similarity or variability changes with separation. Two samples ten metres apart may carry different information from two samples ten kilometres apart, depending on the process and scale.

Spatial dependence is not guaranteed. Nearby values can be related, but measurement error, abrupt boundaries, mixing or a process operating at another scale can weaken that relationship. Geostatistics begins by testing and modelling structure rather than assuming that everything nearby is alike.

Keep the variable, location and observation support together. A concentration measured from a small soil core, an average over a field plot and a satellite pixel are different spatial observations even when they share units. A model built for one support cannot be transferred silently to another. State what each value represents in space.

Spatial autocorrelation can be positive, negative or negligible at a given scale. Positive association means nearby values tend to resemble one another; negative association means neighbours tend to contrast. The same process can show different behaviour across distance bands. Describe the empirical pattern rather than reducing every spatial dataset to “near things are similar.”

A spatial analysis needs a target: estimate a variable at points, calculate an area average, map the chance of exceeding a criterion, or compare patterns through time. Those are different tasks. A single attractive interpolation surface cannot answer all of them without additional assumptions.

Define the domain where inference is intended. Boundaries can follow geology, catchments, administrative areas or a study footprint. Extrapolating beyond sampled conditions is more uncertain than interpolating among observations and may be unsupported. Masking a map outside the intended domain prevents colour from implying knowledge where none exists.

Coordinates must use a known reference system, units and transformation. Latitude and longitude in degrees should not be treated as planar metres in a distance calculation. Duplicate coordinates, swapped axes or mixed datums can produce plausible but false spatial patterns. The coordinate audit comes before the variogram.

Distance itself needs a definition. Straight-line distance may fit a continuous field, while river distance, travel distance or connectivity along a geological layer may better represent another process. Choosing Euclidean distance for convenience can join locations that the process cannot connect. The metric belongs to the conceptual model.

Sampling locations determine which distances, directions and regions are observed. A regular grid covers space systematically; randomisation helps probability-based inference; stratification ensures important zones are represented; nested or multi-scale designs can capture several ranges of variation. Convenience sampling usually overrepresents accessible places.

Clustered observations can dominate descriptive statistics and leave other areas poorly informed. Declustering or design-based weighting may be relevant when estimating a domain distribution, but the remedy depends on why the clustering occurred. A cluster around a suspected hotspot is useful for local detail and poor evidence for the average of an entire region.

Plan quality controls, duplicates and coordinate precision with the design. A laboratory duplicate informs analytical or preparation variability, not spatial variability between locations. A co-located field duplicate includes more of the sampling process. Keep those variance sources distinct so the model does not treat technical repetition as new spatial coverage.

A design also needs the smallest spacing and maximum separation capable of informing the expected range. If every pair is far apart, short-range continuity cannot be estimated. If all points lie in one corner, the model cannot invent credible evidence for the rest of the domain. A pilot survey can guide refinement when the scale is uncertain.

Plot locations and values before fitting a variogram. Inspect histograms, summaries, missingness, detection limits, outliers and relationships with known covariates. A spatial trend can make distant regions systematically different even without the stationary local structure assumed by a simple model.

Skewed variables may motivate a transformation, but transformation changes the scale of modelling and the meaning of back-transformation. An estimate on the log scale cannot be exponentiated naively in every setting without considering bias. Report the model scale and the scale on which predictions are presented.

Outliers deserve investigation, not automatic deletion. They may be errors, local extremes or evidence that the domain contains more than one process. Check identity, coordinates, units, method and field notes. If alternative treatments are plausible, show how they affect the spatial model and final decision.

Detection limits and censored observations require explicit handling. Replacing every non-detect with zero invents measured zeros; replacing it with half a limit can distort a variogram, especially when limits differ. The method should reflect the analytical process, proportion censored and decision question, with sensitivity analysis where choices matter.

A semivariogram summarises how half the squared difference between paired values changes with their separation. Pairs are grouped into distance bins, and their contributions are averaged. Small semivariance at short distance is consistent with nearby similarity; increasing semivariance suggests diminishing spatial association.

The experimental variogram is a data summary, not the final model. Bin width, maximum distance, direction and uneven pair counts affect its appearance. A smooth-looking curve based on few pairs is weaker evidence than a noisier bin supported by many well-distributed pairs.

Plot pair counts and examine influential pairs. If a few extreme differences dominate one bin, the apparent structure may not be stable. A trend can also inflate semivariance with distance. Exploratory maps and trend modelling therefore belong beside the variogram rather than after it.

For two fictional observations of 8 and 12 units, their difference is −4, its square is 16 and the semivariance contribution is half of that, 8 square-units. A variogram bin averages such contributions for pairs within its distance and direction rule. The worked arithmetic is simple; choosing representative pairs and a defensible model is the difficult part.

A fitted variogram model is often described through nugget, range and sill. The nugget represents discontinuity near zero distance and may combine measurement error with variation below the sampling scale. It is not automatically all laboratory error.

The range is the distance beyond which the fitted model reaches or approaches its plateau under the chosen form. It describes the modelled scale of spatial dependence, not a physical boundary where correlation suddenly becomes impossible. The sill relates to the variance level reached by the model.

Several model forms can fit limited data similarly while producing different predictions. Parameter values should be assessed with the sampling geometry, process knowledge and validation. Decimal precision from optimisation does not remove structural uncertainty about model choice.

A nugget-dominated variogram can mean genuine fine-scale variability, substantial measurement error, inadequate sampling spacing or a combination. It should not be interpreted automatically as proof that the variable is random everywhere. New close pairs or better measurement controls can help distinguish those explanations.

Spatial behaviour can differ by direction. Geological layers, prevailing flow or elongated landforms may produce longer continuity along one direction than another. Directional variograms compare pairs within angular tolerances to look for this anisotropy.

Geometric anisotropy changes the apparent range by direction while retaining a similar sill; zonal patterns can be more complex. Before assigning anisotropy, check whether sampling density itself differs by direction. A narrow survey corridor can make directional evidence weak.

A large-scale trend and local spatial dependence should not be confused. Universal or regression kriging can model a trend while treating residual dependence separately. Covariates must be available and meaningful across the prediction domain. A flexible trend fitted only to observed locations can absorb local structure and fail elsewhere.

Directional analysis should use enough pairs in each lag and direction. Narrow angular tolerances improve directional specificity but reduce pair counts; wide tolerances mix directions. Show the tolerance, lag geometry and support rather than reporting an anisotropy ratio without the evidence that produced it.

Consider a fictional one-dimensional transect. Measurements are 8 units at position 0 metres, 12 units at 10 metres and 18 units at 30 metres. We want an illustrative estimate at 20 metres. Suppose a previously fitted ordinary-kriging model produces weights 0.10, 0.60 and 0.30 for these observations. The weights sum to one.

The estimate is 0.10×8 + 0.60×12 + 0.30×18 = 13.4 units. The closest observation receives the largest weight here, but kriging weights are not determined by distance alone. They reflect the fitted covariance or variogram and the spatial relationships among all observations.

This calculation does not prove that 13.4 is the true value at 20 metres. A complete result also reports prediction uncertainty and the model conditions. The example intentionally supplies weights so learners can audit the arithmetic without pretending that three points are enough to fit a reliable real-world variogram.

If the same weights are applied after observation C changes from 18 to 22, the estimate becomes 14.6. This sensitivity calculation shows that C contributes thirty percent of the change. In a real kriging system, changing an observation can also influence fitted model parameters when the model is refitted, so fixed-weight sensitivity and full reanalysis are different exercises.

Kriging is a family of linear prediction methods that uses a fitted spatial dependence model. Ordinary kriging assumes an unknown constant mean within the local model context; simple kriging assumes a known mean; universal or regression approaches represent a trend. The names encode assumptions, not quality rankings.

Under its model, ordinary kriging chooses weights to provide an unbiased linear estimate with minimum prediction variance. The estimate is often smoother than the underlying process because prediction averages information. Smoothness should not be mistaken for the absence of local extremes.

Kriging variance reflects the spatial model and sampling configuration, not every source of uncertainty. It does not automatically include coordinate error, uncertain boundaries, biased sampling, model-selection uncertainty or a wrong conceptual domain. Report what the uncertainty measure contains and what it leaves out.

Cross-validation temporarily removes observations and predicts them from the remaining data. Residuals can reveal bias, dispersion problems and locations where the model performs poorly. Summary measures such as mean error and root mean squared error should be paired with residual plots and spatial inspection.

Leave-one-out cross-validation can be optimistic when nearby clustered observations remain after one point is removed. Spatial block validation withholds regions and can better test transfer to unsampled areas, although it changes the prediction task. Choose a validation design that resembles the intended use.

Compare plausible models using more than one score. A lower average error may conceal severe local failure or poorly calibrated uncertainty. Check standardised residuals, coverage of prediction intervals and performance across subregions. Validation can reject an inadequate model; it cannot prove that one model is uniquely true.

A useful diagnostic maps residuals at their locations. Clusters of positive residuals suggest systematic underprediction in a region; clusters of negative residuals suggest overprediction. A small overall mean can conceal both because opposite regional errors cancel. Spatial validation must therefore retain geography, not only a single score.

Validation metrics should be reported in the variable’s units where possible so error is interpretable. Standardised metrics help compare uncertainty calibration, while unit-based errors show practical magnitude. If values were transformed, evaluate performance on both modelling and decision scales and explain how predictions returned to the original units.

An estimate map shows the central prediction under a model. An uncertainty map shows where evidence is weaker or the model predicts more variability. They answer different questions and should be read together. A decision-maker needs to see both, especially near a threshold or boundary.

Colour choices can exaggerate small differences or hide important ones. Use units, a meaningful legend and a stated prediction support. Avoid unclassed gradients that imply detail finer than the sampling and model justify. Sampling points help readers see which features are observed and which are interpolated.

Probability maps can express the modelled chance that a variable exceeds a specified value. The threshold must have a legitimate source and the probability depends on distributional assumptions. Do not turn a modelled probability into a measured fact or omit the consequences of false positive and false negative decisions.

Uncertainty can guide new sampling. Locations with high prediction variance may be candidates, but the best next sample also depends on access, cost, decision importance and whether uncertainty comes from sparse data or a wrong domain model. Sampling solely at the maximum variance may ignore a boundary or hotspot where a smaller uncertainty still has larger consequences.

Kriging usually produces a smoothed best linear prediction. Conditional geostatistical simulation instead generates multiple spatial realisations consistent with observations and a fitted model. Individual realisations are not forecasts of exactly what the world looks like; together they represent possible spatial patterns under the model.

Simulation is useful when a non-linear quantity depends on connected extremes, volumes above a criterion or spatial continuity. Calculating the target across many realisations can show a distribution rather than one smoothed number.

Realisation count, model assumptions and conditioning data affect results. Check that simulations reproduce relevant marginal and spatial statistics, and distinguish simulation variability from uncertainty sources excluded by the model. A colourful ensemble does not repair a biased sampling design.

When presenting simulations, summarise the decision quantity across realisations: for example, the proportion of a fictional domain above a threshold. Do not choose one dramatic realisation as the map of what will happen. An ensemble communicates a distribution of possible patterns under the model, not a gallery from which to select a preferred story.

Physical or administrative boundaries can interrupt spatial dependence. Predicting across a river, fault or management boundary may be inappropriate if the process changes there. Domains can be modelled separately or with structures that represent the boundary, provided enough evidence exists.

Block prediction estimates an average over an area or volume rather than a value at a point. Its uncertainty and variability differ from point prediction. Comparing a point measurement directly with a block estimate can create an apparent discrepancy that is really a support mismatch.

Aggregation also changes patterns. Fine-scale hotspots can disappear in a large average, while a coarse map can create artificial uniformity. This is related to the modifiable areal unit problem in mapped data. State the resolution at which the question, data and decision operate.

A coastline, fault or soil boundary drawn on a map also has uncertainty. Treating it as an exact wall can overstate knowledge. If the boundary position or permeability is uncertain, compare plausible versions and observe how predictions change. Boundary sensitivity is a model question, not merely a cartographic detail.

Secondary variables such as elevation, geophysics, remote sensing or soil type can improve prediction when they are relevant and available across the domain. Regression kriging models a relationship with covariates and spatial dependence in residuals. Cokriging represents cross-dependence among variables.

A strong correlation at sampled points may not persist elsewhere or at another scale. Covariates can share a trend without a direct causal relationship. Validate the combined model spatially and examine whether improvement occurs across the domain or only where sampling is dense.

Measurement supports must be reconciled. An airborne footprint, a laboratory core and a mapped polygon do not observe the same volume. Resampling them onto one grid does not erase that difference. The Geophysics and Geochemistry roots develop the evidence behind common covariates.

Spatiotemporal data add repeated observations and change through time. Measurements close in space but far apart in time may be less comparable than a purely spatial model assumes. Seasonal cycles, trends and interventions can create temporal structure.

A repeated measurement at one location is not a new spatial location. Separate temporal replication from spatial coverage. If monitoring stations change, apparent trends can reflect network design as well as the process. Preserve station history, methods and detection limits.

Spatiotemporal covariance models can represent dependence across both dimensions, while simpler approaches may model time slices or residuals. The choice follows the question and data density. Do not average away a meaningful event or imply continuous change between widely separated surveys without evidence.

A monitoring network can adapt after each survey, but adaptive placement changes the design. New sites may concentrate where earlier values were high, making later raw averages incomparable with earlier ones. Preserve inclusion probabilities or use a model that accounts for the evolving network if estimating a domain-wide trend.

Preserve raw observations, coordinate reference information, units, sampling support, quality flags and exclusions. Record code, software versions, bin settings, variogram form, parameters, neighbourhood rules and prediction grid. A final raster without its model and provenance is not a reproducible analysis.

Separate observed values from predictions in tables and maps. Label transformations and back-transformations. Where several plausible models influence the result, include sensitivity analysis or an ensemble account instead of choosing silently.

A responsible handover states the domain, target, assumptions, validation, uncertainty and decisions the model can support. It also names conditions that would require new sampling or refitting. The Research, Data Literacy and Critical Thinking roots extend this evidence chain.

Archive prediction grids with their coordinate system, cell size, extent and no-data code. A raster detached from these properties can shift relative to other layers or convert missing areas into zeros. Reproducibility includes spatial metadata as well as statistical code.

Begin with a line of three sample locations rather than software. Ask whether the nearest observation should always receive all the weight and why two clustered observations do not equal two independent directions of evidence. Learners can then see that geometry and dependence both matter.

Use the worked weighted estimate for guided arithmetic. Change the supplied weights while keeping their sum at one, and ask how the estimate changes. Then explain why arbitrary weights are not kriging weights: the spatial model and configuration must produce them. This separates calculation from model construction.

Connect the lesson to Statistics, Probability, Geography and Cartography. The immutable Secondary 1 mathematics learning approach anchors units, weighting and guided-to-independent reasoning; the parent education root anchors transfer.

A good teaching assessment provides a new spatial arrangement rather than only new numbers. Learners should notice that moving two samples closer together changes redundancy even if their values stay the same. That recognition shows transfer from weighted arithmetic to spatial evidence.

Is geostatistics just mapping?

No. It models spatial dependence, prediction and uncertainty; a map is one possible output.

Does nearby always mean similar?

No. That relationship must be examined for the variable, scale, direction and domain.

What is a variogram?

It summarises how paired differences change with distance and sometimes direction, then supports a fitted spatial model.

Is kriging always better than simpler interpolation?

No. Its value depends on a defensible spatial model, sampling and the prediction task.

Why can kriging smooth extremes?

It forms weighted predictions from surrounding evidence, so the prediction surface can be less variable than the underlying process.

Does kriging variance contain all uncertainty?

No. It is conditional on the fitted model and sampling configuration and may omit several design and model uncertainties.

Why show sampling points on a prediction map?

They help readers distinguish measured locations from interpolated areas and see where evidence is sparse.

What makes a geostatistical conclusion strong?

A clear target and support, correct coordinates, representative sampling, justified dependence model, spatial validation and uncertainty communicated with the estimate.